A slow dendritic memory that survives timing jitter

What the work claims

Cortical pyramidal neurons produce regenerative events in their dendrites, driven largely by NMDA receptors, that last many tens of milliseconds. A somatic sodium spike lasts a couple of milliseconds. This paper, by Burger, Rule and O'Leary, asks what that order-of-magnitude gap is for, and answers that the long dendritic plateau equips each dendritic branch with a short, resettable memory of its recent inputs.¹ Because the plateau holds depolarisation open for tens of milliseconds, inputs that arrive at scattered times can still be summed to threshold, and the neuron can fire reliably using as little as a single spike, even when input timing is noisy.

This is a computational and theoretical study, not a tissue experiment. Its evidence is an abstract model, benchmarked against a detailed multi-compartment biophysical model and against a published intracellular recording of an NMDA spike. The claim is bold because it inverts the usual intuition that slow dendritic events sit awkwardly with fast signalling, and argues instead that slowness enables the fastest reliable network computations. The eLife assessment rated the significance important and the strength of evidence compelling, while keeping the finding squarely in the domain of modelling.²

How it works

Start with the biophysics. An NMDA spike is a local, regenerative depolarisation in a dendritic branch: if synaptic drive to that branch is strong enough it crosses a threshold, the response becomes super-linear, and then it saturates into a plateau that stays depolarised for tens of milliseconds before decaying. The authors distil this into what they call a Leaky Integrate and Hold unit, or LIH. A dendritic compartment integrates its inputs; when it crosses threshold it latches into a held depolarised state for a fixed duration, then leaks back to rest. The threshold and the plateau duration are the only two ingredients that matter. Multiple such dendritic compartments couple passively into a somatic compartment, which is a standard leaky integrate and fire unit with a membrane time constant of about 10 milliseconds.

The contrast is stark. In a high-conductance state, the kind a cortical neuron actually sits in during active processing, the effective membrane time constant is short, so excitatory postsynaptic potentials decay within two to three milliseconds. Inputs that do not arrive nearly simultaneously never sum to threshold, and passive summation puts an unrealistic demand on millisecond spike-timing precision. Add the dendritic hold, and asynchronous inputs across branches summate as though they had arrived together. The authors quantified the failure mode by drawing input times from a normal distribution and widening its standard deviation in units of the membrane time constant: without plateaus, jitter both lowers the mean depolarisation and inflates its coefficient of variation, so the cell drops spikes it should have fired. Plateaus of roughly 20 milliseconds filter that variability out. They also checked robustness to inhibition using the detailed biophysical model, and found plateaus survive tonic inhibitory conductance up to roughly the point where total conductance is balanced, with plateau duration shrinking linearly as inhibitory conductance rises. Finally they showed a small network using these units solves an association and discrimination task with sparse spiking under imposed timing jitter.

Where a skeptic should push

The most load-bearing assumption is the combination of a fast, near-binary operating regime, in which each neuron contributes at most one spike per computation, with a plateau that is treated as an approximately fixed-duration hold. The paper's own inhibition result quietly undercuts the second half. If plateau duration falls linearly with inhibitory conductance, then the hold is not a fixed 20 millisecond buffer but a quantity that inhibition can set, lengthen or truncate. The authors are explicit that they modelled only tonic inhibition and did not attempt temporal variation in inhibition, which in a real cortical circuit is exactly what would clip a plateau mid-hold. The jitter tolerance that the whole argument rests on is therefore demonstrated in the regime most favourable to it, with inhibition held still.

The rest is the honest gap between a model and a mechanism. What is demonstrated is that an LIH element makes asynchronous summation and jitter-tolerant firing work in simulation, and that the abstract unit tracks a detailed biophysical model. What is asserted, and framed as a testable hypothesis rather than a result, is that cortex uses plateaus this way to make rapid, sparse decisions. The one-spike-per-computation regime has empirical support in some sensory settings but is an interpretive choice, not a general law. None of this is a flaw in the paper, which is careful; it is a caution against reading a clean in-silico result as an established property of tissue.

What a spike-only readout cannot see

The sharpest consequence for computing on living neural tissue is an observability problem. A microelectrode array reads extracellular signatures of somatic and axonal spikes. If a real and load-bearing part of a neuron's computation is the dendritic hold state, the decisive variable is a branch-local depolarisation that reaches the soma only as summed drive, not as a signal the extracellular electrode records directly. Two organoids could present nearly indistinguishable spike rasters while sitting in different dendritic hold states, one poised to integrate an incoming volley and one about to let it decay. The readout under-observes the integrating variable rather than capturing it, because the mapping from hold state to spike output is many-to-one in the behaviourally relevant regime. The hold state is not strictly unrecoverable, since it leaves a statistical fingerprint in firing reliability and in timing under repeated or paired stimulation, but it cannot be read off directly the way a spike can. This does not overturn spike-based decoding; it bounds it, because a decoder trained on spikes is fitting the visible residue of a slower analog process it cannot observe outright.

The opportunity runs the other way, and it is genuine. Organoid networks are notorious for imprecise, drifting spike timing; asynchrony is their default condition, and it is a large part of why their activity reads as noise. The LIH mechanism says that a neuron endowed with long dendritic plateaus can integrate robustly precisely because inputs do not need to be synchronous. To the extent that maturing cortical organoid neurons express NMDA receptors and develop pyramidal-like dendritic arbors, the mechanism is at least available in principle, though that is a property of tissue the source did not study. If it is present, it offers a reason why untrained tissue might perform reliable integration without the precise wiring and timing that engineered systems require. That reading carries one condition the model did not test: the benefit assumes plateaus are not dynamically truncated by inhibition, and immature organoid inhibition, which is often weak or depolarising early in development, could behave very differently from the tonic inhibition the study modelled. The paper also hands a concrete analog-hardware target as a side effect, a hold element with a roughly 20 millisecond plateau duration and a threshold, which any neuromorphic substrate could copy.

That last point is also the threat, and it is twofold. First, the benefit is maturity and morphology gated. Dendritic plateaus depend on distal, NMDA-rich branches and specific channel expression; a young organoid with immature, electrically compact neurons may simply lack them, so the robustness cannot be assumed to be present in a three-month culture and would have to be shown. Second, the paper's central move is to reproduce the advantage with a two-parameter abstract unit, and the plateau it abstracts is experimentally documented in mammalian cortical pyramidal neurons, not merely a modelling convenience. If a leaky integrate and hold circuit captures the computational benefit, then the primitive is cheap to emulate in silicon, which is an obsolescence argument against paying the metabolic and logistical cost of wet tissue to obtain it. The correct reading is symmetric: this is a hypothesis about what tissue might do and a blueprint for what silicon can copy, and nothing here shows an organoid performing the task.

The bottom line

Established, in a model that tracks a detailed biophysical simulation: a leaky integrate and hold dendritic element lets asynchronous inputs summate and rescues reliable, sparse spiking under timing jitter, where a fast passive soma fails. Still a hypothesis: that cortical neurons, or a brain organoid, actually use plateaus this way for rapid decisions, and that the hold survives the dynamic inhibition the study did not model. What would confirm it is dendritic recording or voltage imaging in tissue showing plateau-gated somatic firing that persists under imposed input jitter and collapses when NMDA plateaus are blocked. What would break it is dynamic inhibition truncating plateaus enough to abolish the jitter tolerance, or the one-spike regime failing to generalise. For anyone building a computer out of neurons, the durable lesson is that the interesting state may live where the electrodes cannot see it. Readers tracking this thread can follow the wider analysis stream for related work on readout and substrate.